A YOLOX-Based Detection Method of Triple-Cascade Feature Level Fusion for Power System External Defects

Abstract

With the digital transformation of the power system, it is of great significance to realize intelligent identification of external breaking defects of overhead transmission lines and power pylons. This paper proposes a YOLOX-based detection method of triple-cascade feature level fusion for power system external defects. Based on YOLOX, the triple-cascade feature level fusion defect recognition and detection method, which is classified layer by layer according to the device inclusion relationship, are adopted. First, the types of equipment are judged, and then the grading standard are determined. Further, the part of the defect types, which are difficult to distinguish in the traditional machine learning algorithm, are refined and identified for the details. Finally, the proposed method is verified based on Python and NVIDIA Jetson TX2 platform with using the image data-set of the overhead transmission lines and power pylons. The mAP value of the model reaches 95.34%, which is higher 10.13% than that of YOLOX, and the detection speed reaches 32fps, which shows a promising performance for the robustness and real-time requirements of defect identification in the new power system.